Learning Facial Action Units from Web Images with Scalable Weakly Supervised Clustering

A Python implementation

Implements train_toy.py where the toy dataset and makemoons (sklearn) generated datasets can be runned. Also now support mnist (mnist package installation required) training and olivetti faces from public dataset clustering with 40 clusters.

Requirements

• scipy
• numpy
• python 3.*
• sklearn
• mnist (optional for train_mnist.py)

Running the repo

Run either of these files

• train_toy.py (default)
• train_mnist.py (mnist package required)
• train_olivetti_faces.py (faces database - https://scikit-learn.org/stable/auto_examples/decomposition/plot_faces_decomposition.html)

Result (on toy dataset)

References

Article

https://openaccess.thecvf.com/content_cvpr_2018/papers/Zhao_Learning_Facial_Action_CVPR_2018_paper.pdf

Original implementation in Matlab:

https://github.com/zkl20061823/WSC

Bonus:

Adam Optimizer added, with amsgrad (default to True)